elem *cilkred_map::insert_no_rehash(__cilkrts_worker *w,
void *key,
__cilkrts_hyperobject_base *hb,
void *val)
{
#if REDPAR_DEBUG >= 2
fprintf(stderr, "[W=%d, desc=insert_no_rehash, this_map=%p]\n",
w->self, this);
verify_current_wkr(w);
#endif
CILK_ASSERT((w == 0 && g == 0) || w->g == g);
CILK_ASSERT(key != 0);
CILK_ASSERT(val != 0);
elem *el = grow(w, &(buckets[hashfun(this, key)]));
#if REDPAR_DEBUG >= 3
fprintf(stderr, "[W=%d, this=%p, inserting key=%p, val=%p, el = %p]\n",
w->self, this, key, val, el);
#endif
el->key = key;
el->hb = hb;
el->val = val;
++nelem;
return el;
}
void cilkred_map::rehash(__cilkrts_worker *w)
{
#if REDPAR_DEBUG >= 1
fprintf(stderr, "[W=%d, desc=rehash, this_map=%p, g=%p, w->g=%p]\n",
w->self, this, g, w->g);
verify_current_wkr(w);
#endif
CILK_ASSERT((w == 0 && g == 0) || w->g == g);
size_t onbuckets = nbuckets;
size_t onelem = nelem;
bucket **obuckets = buckets;
size_t i;
bucket *b;
make_buckets(w, nextsz(nbuckets));
for (i = 0; i < onbuckets; ++i) {
b = obuckets[i];
if (b) {
elem *oel;
for (oel = b->el; oel->key; ++oel)
insert_no_rehash(w, oel->key, oel->hb, oel->val);
}
}
CILK_ASSERT(nelem == onelem);
free_buckets(w, obuckets, onbuckets);
}
struct os_mutex *__cilkrts_os_mutex_create(void)
{
int status;
struct os_mutex *mutex = (struct os_mutex *)malloc(sizeof(struct os_mutex));
pthread_mutexattr_t attr;
ITT_SYNC_CREATE(mutex, "OS Mutex");
if (!mutex) {
if (static_mutex_used) {
__cilkrts_bug("Cilk RTS library initialization failed");
} else {
static_mutex_used = 1;
mutex = &static_mutex;
}
}
status = pthread_mutexattr_init(&attr);
CILK_ASSERT (status == 0);
#if defined DEBUG || CILK_LIB_DEBUG
#ifdef PTHREAD_MUTEX_ERRORCHECK
status = pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_ERRORCHECK);
#else
status = pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_ERRORCHECK_NP);
#endif
CILK_ASSERT (status == 0);
#endif
status = pthread_mutex_init (&mutex->mutex, &attr);
CILK_ASSERT (status == 0);
pthread_mutexattr_destroy(&attr);
return mutex;
}
static Closure *deque_xtract_bottom(CilkWorkerState *const ws, int pn)
{
Closure *cl;
deque_assert_ownership(ws, pn);
cl = USE_PARAMETER(deques)[pn].bottom;
if (cl) {
CILK_ASSERT(ws, cl->owner_ready_deque == pn);
USE_PARAMETER(deques)[pn].bottom = cl->prev_ready;
if (cl == USE_PARAMETER(deques)[pn].top) {
CILK_ASSERT(ws, cl->prev_ready == (Closure *) NULL);
USE_PARAMETER(deques)[pn].top = (Closure *) NULL;
} else {
CILK_ASSERT(ws, cl->prev_ready);
(cl->prev_ready)->next_ready = (Closure *) NULL;
}
WHEN_CILK_DEBUG(cl->owner_ready_deque = NOBODY);
} else {
CILK_ASSERT(ws, USE_PARAMETER(deques)[pn].top == (Closure *) NULL);
}
return cl;
}
void cilk_fiber::reset_state(cilk_fiber_proc start_proc)
{
// Setup the fiber and return.
this->m_start_proc = start_proc;
CILK_ASSERT(!this->is_resumable());
CILK_ASSERT(NULL == this->m_pending_remove_ref);
CILK_ASSERT(NULL == this->m_pending_pool);
}
/**
* Record data for a successful steal.
*
* The pedigree for a STEAL record is the pedigree of the stolen frame.
*
* @note It's assumed that replay_record_steal() has already checked that we're
* recording a log and that the record/replay functionality has not been
* compiled out.
*
* @param w The worker stealing a frame.
* @param victim_id The ID of the worker which had it's frame stolen.
*/
void replay_record_steal_internal(__cilkrts_worker *w, int32_t victim_id)
{
// Follow the pedigree chain using worker's stack frame
CILK_ASSERT(w->l->next_frame_ff);
CILK_ASSERT(w->l->next_frame_ff->call_stack);
// Record steal: STEAL pedigree victim_id thief_id
write_to_replay_log (w, PED_TYPE_STR_STEAL,
&(w->l->next_frame_ff->call_stack->parent_pedigree),
victim_id);
}
void cilk_fiber::deallocate_self(cilk_fiber_pool* pool)
{
this->set_resumable(false);
CILK_ASSERT(NULL != pool);
CILK_ASSERT(!this->is_allocated_from_thread());
this->assert_ref_count_equals(0);
// Cases:
//
// 1. pool has space: Add to this pool.
// 2. pool is full: Give some fibers to parent, and then free
// enough to make space for the fiber we are deallocating.
// Then put the fiber back into the pool.
const bool need_lock = pool->lock;
// Grab the lock for the remaining cases.
if (need_lock) {
spin_mutex_lock(pool->lock);
}
// Case 1: this pool has space. Return the fiber.
if (pool->size < pool->max_size)
{
// Add this fiber to pool
pool->fibers[pool->size++] = this;
if (need_lock) {
spin_mutex_unlock(pool->lock);
}
return;
}
// Case 2: Pool is full.
//
// First free up some space by giving fibers to the parent.
if (pool->parent)
{
// Pool is full. Move all but "num_to_keep" fibers to parent,
// if we can.
unsigned num_to_keep = pool->max_size/2 + pool->max_size/4;
cilk_fiber_pool_move_fibers_to_parent_pool(pool, num_to_keep);
}
if (need_lock) {
spin_mutex_unlock(pool->lock);
}
// Now, free a fiber to make room for the one we need to put back,
// and then put this fiber back. This step may actually return
// fibers to the heap.
cilk_fiber_pool_free_fibers_from_pool(pool, pool->max_size -1, this);
}
CILK_EXPORT
void __cilkrts_hyper_create(__cilkrts_hyperobject_base *hb)
{
// This function registers the specified hyperobject in the current
// reducer map and registers the initial value of the hyperobject as the
// leftmost view of the reducer.
__cilkrts_worker *w = __cilkrts_get_tls_worker();
if (! w) {
// If there is no worker, then there is nothing to do: The iniitial
// value will automatically be used as the left-most view when we
// enter Cilk.
return;
}
// Disable Cilkscreen for the duration of this call. The destructor for
// this class will re-enable Cilkscreen when the method returns. This
// will prevent Cilkscreen from reporting apparent races in reducers
DisableCilkscreen x;
void* key = get_hyperobject_key(hb);
void* view = get_leftmost_view(key);
cilkred_map *h = w->reducer_map;
if (__builtin_expect(!h, 0)) {
h = install_new_reducer_map(w);
#if REDPAR_DEBUG >= 2
fprintf(stderr, "[W=%d, hb=%p, hyper_create, isntalled new map %p, view=%p]\n",
w->self, hb, h, view);
#endif
}
/* Must not exist. */
CILK_ASSERT(h->lookup(key) == NULL);
#if REDPAR_DEBUG >= 3
verify_current_wkr(w);
fprintf(stderr, "[W=%d, hb=%p, lookup in map %p of view %p, should be null]\n",
w->self, hb, h, view);
fprintf(stderr, "W=%d, h=%p, inserting key %p, view%p\n",
w->self,
h,
&(hb->__c_monoid),
view);
#endif
if (h->merging)
__cilkrts_bug("User error: hyperobject used by another hyperobject");
CILK_ASSERT(w->reducer_map == h);
// The address of the leftmost value is the same as the key for lookup.
(void) h->rehash_and_insert(w, view, hb, view);
}
static Closure *deque_peek_bottom(CilkWorkerState *const ws, int pn)
{
Closure *cl;
deque_assert_ownership(ws, pn);
cl = USE_PARAMETER(deques)[pn].bottom;
if (cl) {
CILK_ASSERT(ws, cl->owner_ready_deque == pn);
} else {
CILK_ASSERT(ws, USE_PARAMETER(deques)[pn].top == (Closure *) NULL);
}
return cl;
}
void cilkred_map::make_buckets(__cilkrts_worker *w,
size_t new_nbuckets)
{
nbuckets = new_nbuckets;
CILK_ASSERT(is_power_of_2(nbuckets));
#if defined __GNUC__ && defined __ICC
/* bug workaround -- suppress calls to _intel_fast_memset */
bucket *volatile*new_buckets = (bucket *volatile*)
#else
bucket **new_buckets = (bucket **)
#endif
__cilkrts_frame_malloc(w, nbuckets * sizeof(*(buckets)));
#if REDPAR_DEBUG >= 1
fprintf(stderr, "W=%d, desc=make_buckets, new_buckets=%p, new_nbuckets=%zd\n",
w->self, new_buckets, new_nbuckets);
#endif
for (size_t i = 0; i < new_nbuckets; ++i)
new_buckets[i] = 0;
#if defined __GNUC__ && defined __ICC
buckets = (bucket **)new_buckets;
#else
buckets = new_buckets;
#endif
nelem = 0;
}
void __cilkrts_save_exception_state(__cilkrts_worker *w, full_frame *ff)
{
save_exception_info(w, __cxa_get_globals(), 0, false, "undo-detach");
CILK_ASSERT(NULL == ff->pending_exception);
ff->pending_exception = w->l->pending_exception;
w->l->pending_exception = NULL;
}
NORETURN
cilk_fiber::remove_reference_from_self_and_resume_other(cilk_fiber_pool* self_pool,
cilk_fiber* other)
{
// Decrement my reference count once (to suspend)
// Increment other's count (to resume)
// Suspended fiber should have a reference count of at least 1. (It is not in a pool).
this->dec_ref_count();
other->inc_ref_count();
// Set a pending remove reference for this fiber, once we have
// actually switched off.
other->m_pending_remove_ref = this;
other->m_pending_pool = self_pool;
// Pass along my owner.
other->owner = this->owner;
this->owner = NULL;
// Since we are deallocating self, this fiber does not become
// resumable.
CILK_ASSERT(!this->is_resumable());
cilk_fiber_sysdep* self = this->sysdep();
self->jump_to_resume_other_sysdep(other->sysdep());
__cilkrts_bug("Deallocating fiber. We should never come back here.");
std::abort();
}
void cilk_fiber::suspend_self_and_resume_other(cilk_fiber* other)
{
#if FIBER_DEBUG >=1
fprintf(stderr, "suspend_self_and_resume_other: self =%p, other=%p [owner=%p, resume_sf=%p]\n",
this, other, other->owner, other->resume_sf);
#endif
// Decrement my reference count (to suspend)
// Increment other's count (to resume)
// Suspended fiber should have a reference count of at least 1. (It is not in a pool).
this->dec_ref_count();
other->inc_ref_count();
this->assert_ref_count_at_least(1);
// Pass along my owner.
other->owner = this->owner;
this->owner = NULL;
// Change this fiber to resumable.
CILK_ASSERT(!this->is_resumable());
this->set_resumable(true);
// Normally, I'd assert other->is_resumable(). But this flag may
// be false the first time we try to "resume" a fiber.
cilk_fiber_sysdep* self = this->sysdep();
self->suspend_self_and_resume_other_sysdep(other->sysdep());
// HAVE RESUMED EXECUTION
// When we come back here, we should have at least two references:
// one for the fiber being allocated / out of a pool, and one for it being active.
this->assert_ref_count_at_least(2);
}
void cilk_fiber_pool_destroy(cilk_fiber_pool* pool)
{
CILK_ASSERT(cilk_fiber_pool_sanity_check(pool, "pool_destroy"));
// Lock my own pool, if I need to.
if (pool->lock) {
spin_mutex_lock(pool->lock);
}
// Give any remaining fibers to parent pool.
if (pool->parent) {
cilk_fiber_pool_move_fibers_to_parent_pool(pool, 0);
}
// Unlock pool.
if (pool->lock) {
spin_mutex_unlock(pool->lock);
}
// If I have any left in my pool, just free them myself.
// This method may acquire the pool lock.
cilk_fiber_pool_free_fibers_from_pool(pool, 0, NULL);
// Destroy the lock if there is one.
if (pool->lock) {
spin_mutex_destroy(pool->lock);
}
__cilkrts_free(pool->fibers);
}
void cilk_fiber_pool_set_fiber_limit(cilk_fiber_pool* root_pool,
unsigned max_fibers_to_allocate)
{
// Should only set limit on root pool, not children.
CILK_ASSERT(NULL == root_pool->parent);
root_pool->alloc_max = max_fibers_to_allocate;
}
/*
* Do any necessary cleanup for the logs - See record-replay.h for full
* routine header.
*/
void replay_term(global_state_t *g)
{
// Free memory for the record/replay log file name, if we've got one
if (g->record_replay_file_name)
__cilkrts_free(g->record_replay_file_name);
// Per-worker cleanup
for(int i = 0; i < g->total_workers; ++i)
{
__cilkrts_worker *w = g->workers[i];
// Close the log files, if we've opened them
if(w->l->record_replay_fptr)
fclose(w->l->record_replay_fptr);
if (w->l->replay_list_root)
{
// We should have consumed the entire list
CILK_ASSERT(ped_type_last == w->l->replay_list_entry->m_type);
replay_entry_t *entry = w->l->replay_list_root;
while (ped_type_last != entry->m_type)
{
// Free the pedigree memory for each entry
entry->unload();
entry++;
}
__cilkrts_free(w->l->replay_list_root);
w->l->replay_list_root = NULL;
w->l->replay_list_entry = NULL;
}
}
}
/* debugging support: check consistency of a reducer map */
void cilkred_map::check(bool allow_null_val)
{
size_t count = 0;
CILK_ASSERT(buckets);
for (size_t i = 0; i < nbuckets; ++i) {
bucket *b = buckets[i];
if (b)
for (elem *el = b->el; el->key; ++el) {
CILK_ASSERT(allow_null_val || el->val);
++count;
}
}
CILK_ASSERT(nelem == count);
/*global_reducer_map::check();*/
}
void __cilkrts_os_mutex_unlock(struct os_mutex *p)
{
int status;
ITT_SYNC_RELEASING(p);
status = pthread_mutex_unlock (&p->mutex);
CILK_ASSERT(status == 0);
}
__CILKRTS_BEGIN_EXTERN_C
/**
* Walk the pedigree and generate a string representation with underscores
* between terms. Currently does a recursive walk to generate a forward
* pedigree.
*
* @param p The buffer that is to be filled. Assumed to be PEDIGREE_BUFF_SIZE
* characters long
* @param pnode The initial pedigree term to be written.
*
* @return A pointer into the pedigree string buffer after a term has been
* written.
*/
static
char * walk_pedigree_nodes(char *p, const __cilkrts_pedigree *pnode)
{
CILK_ASSERT(pnode);
if (pnode->parent)
{
p = walk_pedigree_nodes(p, pnode->parent);
p += cilk_snprintf_s(p, PEDIGREE_BUFF_SIZE, "%s", (char *)"_");
}
return p + cilk_snprintf_l(p, PEDIGREE_BUFF_SIZE, "%" PRIu64, pnode->rank);
}
/* Destroy a reducer map. The map must have been allocated
from the worker's global context and should have been
allocated from the same worker. */
void __cilkrts_destroy_reducer_map(__cilkrts_worker *w, cilkred_map *h)
{
CILK_ASSERT((w == 0 && h->g == 0) || w->g == h->g);
verify_current_wkr(w);
/* the reducer map is allowed to contain el->val == NULL here (and
only here). We set el->val == NULL only when we know that the
map will be destroyed immediately afterwards. */
DBG h->check(/*allow_null_val=*/true);
bucket *b;
size_t i;
for (i = 0; i < h->nbuckets; ++i) {
b = h->buckets[i];
if (b) {
elem *el;
for (el = b->el; el->key; ++el) {
if (el->val)
el->destroy();
}
}
}
free_buckets(w, h->buckets, h->nbuckets);
#if REDPAR_DEBUG >= 1
fprintf(stderr, "W=%d, destroy_red_map, freeing map h=%p, size=%zd\n",
w->self, h, sizeof(*h));
#endif
__cilkrts_frame_free(w, h, sizeof(*h));
}
void __cilkrts_mutex_lock(__cilkrts_worker *w, struct mutex *m)
{
int count;
const int maxspin = 1000; /* SWAG */
NOTE_INTERVAL(w, INTERVAL_MUTEX_LOCK);
if (!TRY_ACQUIRE(m)) {
START_INTERVAL(w, INTERVAL_MUTEX_LOCK_SPINNING);
count = 0;
do {
do {
__cilkrts_short_pause();
if (++count >= maxspin) {
STOP_INTERVAL(w, INTERVAL_MUTEX_LOCK_SPINNING);
START_INTERVAL(w, INTERVAL_MUTEX_LOCK_YIELDING);
/* let the OS reschedule every once in a while */
__cilkrts_yield();
STOP_INTERVAL(w, INTERVAL_MUTEX_LOCK_YIELDING);
START_INTERVAL(w, INTERVAL_MUTEX_LOCK_SPINNING);
count = 0;
}
} while (m->lock != 0);
} while (!TRY_ACQUIRE(m));
STOP_INTERVAL(w, INTERVAL_MUTEX_LOCK_SPINNING);
}
CILK_ASSERT(m->owner == 0);
m->owner = w;
}
CILK_EXPORT void* __CILKRTS_STRAND_PURE(
__cilkrts_hyper_lookup(__cilkrts_hyperobject_base *hb))
{
__cilkrts_worker* w = __cilkrts_get_tls_worker_fast();
void* key = get_leftmost_view(hb);
if (! w)
return get_leftmost_view(key);
// Disable Cilkscreen for the duration of this call. This will
// prevent Cilkscreen from reporting apparent races in reducers
DisableCilkscreen dguard;
if (__builtin_expect(w->g->force_reduce, 0))
__cilkrts_promote_own_deque(w);
cilkred_map* h = w->reducer_map;
if (__builtin_expect(!h, 0)) {
h = install_new_reducer_map(w);
}
if (h->merging)
__cilkrts_bug("User error: hyperobject used by another hyperobject");
elem* el = h->lookup(key);
if (! el) {
/* lookup failed; insert a new default element */
void *rep;
{
/* re-enable cilkscreen while calling the constructor */
EnableCilkscreen eguard;
if (h->is_leftmost)
{
// This special case is called only if the reducer was not
// registered using __cilkrts_hyper_create, e.g., if this is a
// C reducer in global scope or if there is no bound worker.
rep = get_leftmost_view(key);
}
else
{
rep = hb->__c_monoid.allocate_fn((void*)hb,
hb->__view_size);
// TBD: Handle exception on identity function
hb->__c_monoid.identity_fn((void*)hb, rep);
}
}
#if REDPAR_DEBUG >= 3
fprintf(stderr, "W=%d, h=%p, inserting key %p, val%p\n",
w->self,
h,
&(hb->__c_monoid),
rep);
CILK_ASSERT(w->reducer_map == h);
#endif
el = h->rehash_and_insert(w, key, hb, rep);
}
return el->val;
}
/**
* Advance to the next log entry from a SYNC record. Consume the current
* SYNC record on this worker and advance to the next one.
*
* @note It's assumed that replay_advance_from_sync() has already returned if
* we're not replaying a log, or if record/replay functionality has been
* compiled out.
*
* @param w The worker whose replay log we're advancing.
*/
void replay_advance_from_sync_internal (__cilkrts_worker *w)
{
// The current replay entry must be a SYNC
CILK_ASSERT(ped_type_sync == w->l->replay_list_entry->m_type);
// Advance to the next entry
w->l->replay_list_entry = w->l->replay_list_entry->next_entry();
}
static void deque_add_bottom(CilkWorkerState *const ws, Closure *cl, int pn)
{
deque_assert_ownership(ws, pn);
CILK_ASSERT(ws, cl->owner_ready_deque == NOBODY);
cl->prev_ready = USE_PARAMETER(deques)[pn].bottom;
cl->next_ready = (Closure *)NULL;
USE_PARAMETER(deques)[pn].bottom = cl;
WHEN_CILK_DEBUG(cl->owner_ready_deque = pn);
if (USE_PARAMETER(deques)[pn].top) {
CILK_ASSERT(ws, cl->prev_ready);
(cl->prev_ready)->next_ready = cl;
} else {
USE_PARAMETER(deques)[pn].top = cl;
}
}
void __cilkrts_note_interval(__cilkrts_worker *w, enum interval i)
{
if (w) {
statistics *s = w->l->stats;
CILK_ASSERT(s->start[i] == INVALID_START);
s->count[i]++;
}
}
static inline void Closure_destroy(CilkWorkerState *const ws, Closure *t)
{
Closure_checkmagic(ws, t);
WHEN_CILK_DEBUG(t->magic = ~CILK_CLOSURE_MAGIC);
CILK_ASSERT(ws, ! (t->malloced));
Closure_clean(ws->context, t);
Cilk_internal_free(ws, t, sizeof(Closure));
}
void __cilkrts_stop_interval(__cilkrts_worker *w, enum interval i)
{
if (w) {
statistics *s = w->l->stats;
CILK_ASSERT(s->start[i] != INVALID_START);
s->accum[i] += __cilkrts_getticks() - s->start[i];
s->start[i] = INVALID_START;
}
}
void load_pedigree_leaf_into_user_worker(__cilkrts_worker *w)
{
__cilkrts_pedigree *pedigree_leaf;
CILK_ASSERT(w->l->type == WORKER_USER);
pedigree_leaf = __cilkrts_get_tls_pedigree_leaf(1);
w->pedigree = *pedigree_leaf;
// Save a pointer to the old leaf.
// We'll need to restore it later.
CILK_ASSERT(w->l->original_pedigree_leaf == NULL);
w->l->original_pedigree_leaf = pedigree_leaf;
__cilkrts_set_tls_pedigree_leaf(&w->pedigree);
// Check that this new pedigree root has at least two values.
CILK_ASSERT(w->pedigree.parent);
CILK_ASSERT(w->pedigree.parent->parent == NULL);
}
void global_os_mutex_unlock(void)
{
// We'd better have allocated a global_os_mutex. This means you should
// have called global_os_mutex_lock() before calling
// global_os_mutex_unlock(), but this is the only check for it.
CILK_ASSERT(NULL != global_os_mutex);
// Release the global OS mutex
__cilkrts_os_mutex_unlock(global_os_mutex);
}
static void global_free(global_state_t *g, void *mem, int bucket)
{
size_t size;
CILK_ASSERT(bucket < FRAME_MALLOC_NBUCKETS);
size = FRAME_MALLOC_BUCKET_TO_SIZE(bucket);
g->frame_malloc.allocated_from_global_pool -= size;
push(&g->frame_malloc.global_free_list[bucket], mem);
}
